approximating the normal distribution with GAN

GAN_normal.py

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

mu,sigma=5.0,0.2
latent_dim=2

def mlp(input):
    w1=tf.get_variable("w1", [input.get_shape()[1], 8], initializer=tf.random_normal_initializer())
    b1=tf.get_variable("b1", [8], initializer=tf.constant_initializer(0.1))
    w2=tf.get_variable("w2", [8,8], initializer=tf.random_normal_initializer())
    b2=tf.get_variable("b2", [8], initializer=tf.constant_initializer(0.1))
    w3=tf.get_variable("w3", [8,8], initializer=tf.random_normal_initializer())
    b3=tf.get_variable("b3", [8], initializer=tf.constant_initializer(0.1))
    w4=tf.get_variable("w4", [8,8], initializer=tf.random_normal_initializer())
    b4=tf.get_variable("b4", [8], initializer=tf.constant_initializer(0.0))
    w_=tf.get_variable("w_", [8,1], initializer=tf.random_normal_initializer())
    b_=tf.get_variable("b_", [1], initializer=tf.constant_initializer(0.0))
    fc1=tf.nn.dropout(tf.nn.relu(tf.matmul(input,w1)+b1),0.5)
    fc2=tf.nn.dropout(tf.nn.relu(tf.matmul(fc1,w2)+b2),0.5)
    fc3=tf.nn.dropout(tf.nn.relu(tf.matmul(fc2,w3)+b3),0.5)
    fc4=tf.nn.tanh(tf.matmul(fc3,w4)+b4)
    fc_=tf.matmul(fc4,w_)+b_
    out=fc_
    out_s=tf.sigmoid(fc_)
    return out,out_s,[w1,b1,w2,b2,w3,b3,w4,b4,w_,b_]

def optimizer(loss,var_list):
    return tf.train.AdamOptimizer().minimize(loss,var_list=var_list)

with tf.variable_scope("G"):
    z_node=tf.placeholder(tf.float32, [None,latent_dim])
    G,_,theta_g=mlp(z_node)
with tf.variable_scope("D") as scope:
    x_node=tf.placeholder(tf.float32, [None,1])
    _,D1,theta_d=mlp(x_node)
    scope.reuse_variables()
    _,D2,theta_d=mlp(G)
obj_d=tf.reduce_mean(tf.log(1e-10+D1)+tf.log(1e-10+1-D2))
obj_g=tf.reduce_mean(tf.log(1e-10+1-D2))

opt_d=optimizer(-obj_d,theta_d)
opt_g=optimizer(obj_g,theta_g)

sess=tf.Session()
sess.run(tf.initialize_all_variables())

TRAIN_ITERS = 10000
BATCH_SIZE = 50
histd, histg= np.zeros(TRAIN_ITERS), np.zeros(TRAIN_ITERS)
for i in range(TRAIN_ITERS):
    for j in range(1):
        x=np.random.normal(mu,sigma,BATCH_SIZE)
        z=np.random.random(BATCH_SIZE*latent_dim)
        histd[i],_=sess.run([obj_d,opt_d], {x_node: np.reshape(x,(BATCH_SIZE,1)), z_node: np.reshape(z,(BATCH_SIZE,latent_dim))})
    z=np.random.random(BATCH_SIZE*latent_dim)
    histg[i],_=sess.run([obj_g,opt_g], {z_node: np.reshape(z,(BATCH_SIZE,latent_dim))})
    if i % 100 == 0:
        print i, histd[i], histg[i]

z=np.random.random(10000*latent_dim)
x=sess.run(G, {z_node: np.reshape(z,(10000,latent_dim))})
plt.hist(x, bins=50, normed=True)
plt.show()